ACCELERATING NEUROSTIMULATION PERSONALIZATION BY LEARNING PARAMETER INTERACTIONS

The therapeutic success of neurostimulation treatments depends on selectively targeting neural circuits specific to each individual. Current clinical applications still rely on inefficient trial-and-error calibration protocols to find the best combination of neurostimulation parameters (NPs) — a process that is increasingly infeasible in high-dimensional settings. Gaussian Process Bayesian Optimization (GPBO) has emerged as an efficient solution for real-time optimization of NPs. However, current GPBO approaches ignore the structural separability of stimulation parameters, a feature that clinical protocols exploit by treating spatial and temporal dimensions separately. In this work, we propose SobolGP, a novel BO framework that incorporates global sensitivity analysis during training to explicitly capture these NP interactions. By optimizing how information propagates within the GP from the learned interactions, this framework not only accelerates the optimization process, but also facilitates scientific discovery by validating hypotheses about NPs. We first validated the enhanced performance of SobolGP using a diverse set of high-dimensional black-box optimization functions. Furthermore, we demonstrate that this framework effectively identifies optimal stimulation parameters across animal models and stimulation modalities — we report an 8% exploration score improvement in a non-human primate 2D cortical stimulation task and notably achieve a 2x convergence speedup and a 4% exploration score improvement in a high-dimensional (5D) rat cortical stimulation task. We are currently expanding these results to human subjects and additional neurostimulation modalities, including transcranial magnetic stimulation, spinal electrical stimulation, and peripheral electrode stimulation. These findings suggest that SobolGP provides a scalable and efficient solution for improved personalization of neurostimulation therapies.